Model Catalog
The model catalog is where you browse, download, and manage transcription models. Open it from Settings > Dictation Models. Each model card shows download and disk size, RAM requirement, accuracy and speed ratings, language support, and a word error rate (WER) benchmark where available. For the mechanics of downloading, switching, and deleting, see Downloading & Managing Models.
Whisper Local models
OpenAI's Whisper architecture running locally via Core ML, optimized for Apple Silicon. Multilingual variants support 99 languages with automatic detection; .en variants are English-only but slightly faster and more accurate at the same size.
| Model | Download | RAM | Accuracy | Speed | Languages |
|---|---|---|---|---|---|
| Tiny | 75 MB | ~400 MB | Basic (~22% WER) | Fastest | 99 |
| Tiny (English) | 66 MB | ~350 MB | Basic | Fastest | English |
| Base | 142 MB | ~500 MB | Fair (~15% WER) | Very fast | 99 |
| Base (English) | 145 MB | ~450 MB | Fair | Very fast | English |
| Small | 466 MB | ~1 GB | Good (~12% WER) | Fast | 99 |
| Small (English) | 483 MB | ~950 MB | Good | Fast | English |
| Medium (English) | 1.5 GB | ~2.3 GB | Excellent (~11% WER) | Moderate | English |
| Large V3 Turbo | 1.6 GB | ~3 GB | Excellent (~8% WER) | Moderate | 99 |
| Large V3 | 3.1 GB | ~5 GB | Best (~7% WER) | Slow | 99 |
Base is the default for new installs — a sensible starting point that downloads quickly. Small is the recommended upgrade for most people: a noticeably better balance of speed and accuracy. Large V3 is the most accurate multilingual model Murmur can run on-device.
NVIDIA Parakeet models
NVIDIA's state-of-the-art ASR models, running on-device via Core ML. They top the open ASR accuracy leaderboards while being dramatically faster than comparable Whisper models — Parakeet can process an hour of audio in well under a minute.
| Model | Download | RAM | Accuracy | Languages |
|---|---|---|---|---|
| Parakeet V2 | ~600 MB | ~1.2 GB | Best (6.05% WER) | English only |
| Parakeet V3 | ~600 MB | ~1.2 GB | Best (6.32% WER) | 25 European languages, auto-detected |
Parakeet V2 is the best choice for English dictation, full stop. Parakeet V3 trades a hair of English accuracy for automatic detection across 25 European languages (you cannot pick the language manually — V3 detects it).
Parakeet V2 and V3 share their files on disk: downloading one downloads both, and deleting one deletes both.
Apple Speech
macOS's built-in speech recognition (the SpeechAnalyzer engine on macOS Tahoe). Nothing to download — it is always available the moment Murmur is installed, which also makes it the automatic fallback if you delete your active model.
- Good accuracy (~14% WER benchmark), very fast
- 40+ languages, managed by macOS — you pick a locale such as "English (US)"
- No download, no disk space, no API key
OpenAI cloud models
Cloud transcription using your own OpenAI API key. Audio is sent to OpenAI's servers — review their privacy policy if that matters for your content. Requires internet; files are limited to 25 minutes / 25 MB per request.
| Model | Accuracy | Notes |
|---|---|---|
| GPT-4o Transcribe | Best (~5% WER est.) | Strongest option for noisy or heavily accented audio |
| GPT-4o mini Transcribe | Excellent (~8% WER est.) | Faster and cheaper, great for clean speech |
Both support 99 languages, language hints, and transcription prompts. Set up your API key in Settings > AI Providers.
Which model should you pick?
- English dictation: Parakeet V2 — fastest and most accurate
- Multiple European languages: Parakeet V3
- Languages beyond Europe, on-device: Whisper Small or Large V3 Turbo
- Zero setup / minimal disk: Apple Speech
- Toughest audio, accuracy above all: OpenAI GPT-4o Transcribe (cloud)
- Live streaming text while you speak: a Whisper model — see Engine Selection
Murmur also factors in your Mac's RAM when surfacing recommendations: the Recommended filter in the catalog shows a curated set (Parakeet V2/V3, Apple Speech, Whisper Base and Base English) that runs well on any Apple Silicon Mac.